Luciano Agostini
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Luciano Agostini has a total of 274 co-authors. Here are the top co-authors by number of shared publications:
Marcelo Porto
67 shared publicationsBruno Zatt
66 shared publications- Guilherme Correa
UFPEL
50 shared publications - Marcelo Porto
UFPEL
45 shared publications Sergio Bampi
Informatics Institute (PPGC-PGMICRO), Federal University of Rio Grande do Sul - Porto Alegre, Brazil
38 shared publicationsGustavo Sanchez
IF Farroupilha, Alegrete
34 shared publications
Latest Publications
- Open Access
- 8 October 2026
- Published by Elsevier in Signal Processing: Image Communication
- Vol. 150
Module-wise post-training quantization for a real-time neural video codec: A DCVC-RT case study
- 8 October 2026
- Published by Elsevier in Signal Processing: Image Communication
- Vol. 150
Open AccessModule-wise post-training quantization for a real-time neural video codec: A DCVC-RT case study
- Open Access
- 12 April 2026
- Published by Journal of Integrated Circuits and Systems in Journal of Integrated Circuits and Systems
- Vol. 21 (1) , 1-11
Accelerating Versatile Video Coding via Decision-Tree-Based Fast Affine Motion Estimation
As the demand for video transmission surges in remote work, education, and streaming services, the need for continuous advancements in video encoding technologies becomes increasingly evident. Adapting to the evolving demands of efficient video delivery and consumption requires continuous advancement and refinement of video encoding standards. In this scenario, the Versatile Video Coding (VVC) stands out as a leading example, representing the current state-of-the-art in video coding. This work presents a Machine Learning (ML) approach to accelerate one of the novel tools within the VVC: the Affine Motion Estimation (AME). AME is part of the VVC Inter-Frame Prediction, and it is one of the most computationally intensive tools within VVC. The hardware-friendly Decision Tree ML method was explored in this work to accelerate the AME. A total of 12 Decision Trees were trained to determine whether to skip or not to skip the AME for each block size supported by AME in VVC. The proposed approach achieved an average reduction of 5.54% in the total VVC encoding time and an average reduction of 63.20% in the AME execution time, resulting in an average BD-BR efficiency loss of only 0.50%. These results demonstrate that the strategy of using Machine Learning to reduce the computational cost of VVC has the potential to continue yielding significant results in future solutions.
- 12 April 2026
- Published by Journal of Integrated Circuits and Systems in Journal of Integrated Circuits and Systems
- Vol. 21 (1) , 1-11
Open AccessAccelerating Versatile Video Coding via Decision-Tree-Based Fast Affine Motion Estimation
As the demand for video transmission surges in remote work, education, and streaming services, the need for continuous advancements in video encoding technologies becomes increasingly evident. Adapting to the evolving demands of efficient video delivery and consumption requires continuous advancement and refinement of video encoding standards. In this scenario, the Versatile Video Coding (VVC) stands out as a leading example, representing the current state-of-the-art in video coding. This work presents a Machine Learning (ML) approach to accelerate one of the novel tools within the VVC: the Affine Motion Estimation (AME). AME is part of the VVC Inter-Frame Prediction, and it is one of the most computationally intensive tools within VVC. The hardware-friendly Decision Tree ML method was explored in this work to accelerate the AME. A total of 12 Decision Trees were trained to determine whether to skip or not to skip the AME for each block size supported by AME in VVC. The proposed approach achieved an average reduction of 5.54% in the total VVC encoding time and an average reduction of 63.20% in the AME execution time, resulting in an average BD-BR efficiency loss of only 0.50%. These results demonstrate that the strategy of using Machine Learning to reduce the computational cost of VVC has the potential to continue yielding significant results in future solutions.
- Open Access
- 12 April 2026
- Published by Journal of Integrated Circuits and Systems in Journal of Integrated Circuits and Systems
- Vol. 21 (1) , 1-10
On the Impact of Quantization Precision and Calibration Settings Applied to the SSF Neural Video Decoder
Neural video codecs have recently achieved state-of-the-art performance, but their reliance on floating-point arithmetic introduces cross-platform inconsistencies. In this work, we present an extended investigation of post-training quantization (PTQ) applied to the hyper prior decoding path in the Scale-Space Flow (SSF) codec, focusing on its role in ensuring cross-platform consistency while maintaining compression efficiency. We systematically evaluate the impact of different quantization bit-depths (4-, 8-, 12-, and 16-bit integers), calibration configurations (crop sizes and dataset sizes), and temporal test scenarios (short and long P-frame chains). The results show that aggressive quantization (INT4) severely degrades rate-distortion (RD) performance, whereas INT8 achieves a favorable trade-off with BD-Rate penalties below 7%. Furthermore, INT12 precision is sufficient to match the FP32 baseline with virtually no loss. Calibration analysis reveals that high-precision quantization is only marginally affected by dataset size and resolution, while lower-precision scenarios exhibit less predictable behavior. Finally, an estimation of energy consumption for a dedicated hardware design indicates that quantization can reduce energy consumption by up to 84% compared to the FP32 anchor model, while preserving near-floating-point efficiency.
- 12 April 2026
- Published by Journal of Integrated Circuits and Systems in Journal of Integrated Circuits and Systems
- Vol. 21 (1) , 1-10
Open AccessOn the Impact of Quantization Precision and Calibration Settings Applied to the SSF Neural Video Decoder
Neural video codecs have recently achieved state-of-the-art performance, but their reliance on floating-point arithmetic introduces cross-platform inconsistencies. In this work, we present an extended investigation of post-training quantization (PTQ) applied to the hyper prior decoding path in the Scale-Space Flow (SSF) codec, focusing on its role in ensuring cross-platform consistency while maintaining compression efficiency. We systematically evaluate the impact of different quantization bit-depths (4-, 8-, 12-, and 16-bit integers), calibration configurations (crop sizes and dataset sizes), and temporal test scenarios (short and long P-frame chains). The results show that aggressive quantization (INT4) severely degrades rate-distortion (RD) performance, whereas INT8 achieves a favorable trade-off with BD-Rate penalties below 7%. Furthermore, INT12 precision is sufficient to match the FP32 baseline with virtually no loss. Calibration analysis reveals that high-precision quantization is only marginally affected by dataset size and resolution, while lower-precision scenarios exhibit less predictable behavior. Finally, an estimation of energy consumption for a dedicated hardware design indicates that quantization can reduce energy consumption by up to 84% compared to the FP32 anchor model, while preserving near-floating-point efficiency.
- Open Access
- 3 February 2026
- Published by Springer Nature in Multimedia Tools and Applications
- Vol. 85 (2) , 123
Quantitative assessment of inter-frame prediction in versatile video coding standard
Versatile Video Coding (VVC) is the latest video coding standard established by ISO and ITU-T, boasting a doubled coding efficiency compared to previous standards. However, this substantial enhancement in coding efficiency comes at the cost of an increase in computational demands, posing challenges for practical implementations. This paper conducts a thorough and quantitative assessment of VVC inter-frame prediction, the most computationally intensive tool in the VVC encoder. A comprehensive set of experiments on inter-frame prediction is presented, with a detailed evaluation of the primary bottlenecks of this tool. To the best of the authors’ knowledge, this work is the first in the literature to offer this in-depth assessment of VVC inter-frame prediction.
- 3 February 2026
- Published by Springer Nature in Multimedia Tools and Applications
- Vol. 85 (2) , 123
Open AccessQuantitative assessment of inter-frame prediction in versatile video coding standard
Versatile Video Coding (VVC) is the latest video coding standard established by ISO and ITU-T, boasting a doubled coding efficiency compared to previous standards. However, this substantial enhancement in coding efficiency comes at the cost of an increase in computational demands, posing challenges for practical implementations. This paper conducts a thorough and quantitative assessment of VVC inter-frame prediction, the most computationally intensive tool in the VVC encoder. A comprehensive set of experiments on inter-frame prediction is presented, with a detailed evaluation of the primary bottlenecks of this tool. To the best of the authors’ knowledge, this work is the first in the literature to offer this in-depth assessment of VVC inter-frame prediction.
- Open Access
- 20 January 2026
- Published by Sociedade Brasileira de Computação in Journal of the Brazilian Computer Society
- Vol. 32 (1) , 13-25
A Coding-Efficiency Analysis of HEVC Encoder Embedded in High-End Mobile Chipsets
High-end mobile devices require dedicated hardware for real-time video encoding and decoding processes. However, the inherent complexity of the video encoding process, combined with the physical limitations imposed by hardware design such as energy consumption, encoding time, memory usage, and heat dissipation, demands the implementation of various constraints and limitations in commercial hardware to simplify and make them feasible for general use. The High Efficiency Video Coding (HEVC) standard is the main targeted video encoder for processing high-resolution videos in high-end chipsets. This paper aims to analyze the HEVC encoder implemented into three commercial chipsets found in high-end smartphones (Apple iPhone 14 Pro, Samsung Galaxy S23 Plus, and Redmi Note 10S) from three major mobile chip manufacturers (Apple, Qualcomm, and MediaTek), considering the impacts of video encoder limitations on encoding efficiency (BD-Rate) and encoding time. The results in this paper may be used as a comparative foundation for hardware designers and future works in the field, as it exposes the encoding efficiency drawbacks and the encoding time gains that commercial chipsets exhibit in their HEVC encoder.
- 20 January 2026
- Published by Sociedade Brasileira de Computação in Journal of the Brazilian Computer Society
- Vol. 32 (1) , 13-25
Open AccessA Coding-Efficiency Analysis of HEVC Encoder Embedded in High-End Mobile Chipsets
High-end mobile devices require dedicated hardware for real-time video encoding and decoding processes. However, the inherent complexity of the video encoding process, combined with the physical limitations imposed by hardware design such as energy consumption, encoding time, memory usage, and heat dissipation, demands the implementation of various constraints and limitations in commercial hardware to simplify and make them feasible for general use. The High Efficiency Video Coding (HEVC) standard is the main targeted video encoder for processing high-resolution videos in high-end chipsets. This paper aims to analyze the HEVC encoder implemented into three commercial chipsets found in high-end smartphones (Apple iPhone 14 Pro, Samsung Galaxy S23 Plus, and Redmi Note 10S) from three major mobile chip manufacturers (Apple, Qualcomm, and MediaTek), considering the impacts of video encoder limitations on encoding efficiency (BD-Rate) and encoding time. The results in this paper may be used as a comparative foundation for hardware designers and future works in the field, as it exposes the encoding efficiency drawbacks and the encoding time gains that commercial chipsets exhibit in their HEVC encoder.
- 12 November 2025
- Published by Sociedade Brasileira de Computação
- p. 300-303
Arquitetura de Hardware Dedicada para CNNs: Uma Análise comparativa entre FP32 e INT8
Para viabilizar a implementação de Redes Neurais Convolucionais (CNNs) de alto custo computacional em sistemas restritos, a quantização de modelos, como a conversão de representações de ponto flutuante (float32) para inteiros de 8 bits (int8), é uma técnica essencial. Este trabalho quantifica os ganhos da técnica através da comparação de uma camada convolucional e uma totalmente conectada, descritas em VHDL e sintetizadas para um ASIC de 40 nm. A análise mensura a redução em área e potência da implementação em int8 frente à de float32, validando sua eficácia para hardware de baixo consumo.
- 12 November 2025
- Published by Sociedade Brasileira de Computação
- p. 300-303
Arquitetura de Hardware Dedicada para CNNs: Uma Análise comparativa entre FP32 e INT8
Para viabilizar a implementação de Redes Neurais Convolucionais (CNNs) de alto custo computacional em sistemas restritos, a quantização de modelos, como a conversão de representações de ponto flutuante (float32) para inteiros de 8 bits (int8), é uma técnica essencial. Este trabalho quantifica os ganhos da técnica através da comparação de uma camada convolucional e uma totalmente conectada, descritas em VHDL e sintetizadas para um ASIC de 40 nm. A análise mensura a redução em área e potência da implementação em int8 frente à de float32, validando sua eficácia para hardware de baixo consumo.
- 12 November 2025
- Published by Sociedade Brasileira de Computação
- p. 424-427
CNN Architecture Assessment: Exploring Depth, Width, and Kernel Size for Image Classification
This paper investigates the impact of different convolutional neural network (CNN) architectures on image classification performance using the CIFAR-10 dataset. We evaluate variations in the number of convolutional layers, kernel size, and number of filters. Experiments indicate that increasing the number of filters and slightly enlarging kernel size generally improves accuracy, while deeper models do not always yield better results. The CIFAR-10 dataset contains 60,000 color images across 10 classes, with 45,000 for training, 5,000 for validation, and 10,000 for final testing.
- 12 November 2025
- Published by Sociedade Brasileira de Computação
- p. 424-427
CNN Architecture Assessment: Exploring Depth, Width, and Kernel Size for Image Classification
This paper investigates the impact of different convolutional neural network (CNN) architectures on image classification performance using the CIFAR-10 dataset. We evaluate variations in the number of convolutional layers, kernel size, and number of filters. Experiments indicate that increasing the number of filters and slightly enlarging kernel size generally improves accuracy, while deeper models do not always yield better results. The CIFAR-10 dataset contains 60,000 color images across 10 classes, with 45,000 for training, 5,000 for validation, and 10,000 for final testing.
- Open Access
- 10 November 2025
- Published by Sociedade Brasileira de Computação
- p. 496-500
A Hardware Architecture for Smoothing Filters in H.266/VVC Intra-Prediction
This paper presents a hardware architecture for the preprocessing smoothing filters used in the intra prediction module of the H.266 Versatile Video Coding (H.266/VVC) standard, applicable to both the encoder and decoder sides. The design was fully described in VHDL and synthesized to an Intel Cyclone V FPGA, utilizing 6,961 adaptive logic modules and 4,158 registers, demonstrating feasibility for FPGA implementation with room for integration with a larger intra prediction module. Operating at a maximum frequency of 96.3 MHz without pipelining, the architecture achieves a decoding throughput of up to 2,972 frames per second (fps) for the 1080p resolution, 743 fps for UHD 4K, and 185 fps for UHD 8K, showing that a much lower clock operation can be used for real-time performance. For encoding, the throughput was estimated by targeting a 60 fps rate at 1080p resolution, resulting in an upper bound of approximately 3,171 CUs evaluated per CTU, compatible with fast decision algorithms used in practical encoders.
- 10 November 2025
- Published by Sociedade Brasileira de Computação
- p. 496-500
Open AccessA Hardware Architecture for Smoothing Filters in H.266/VVC Intra-Prediction
This paper presents a hardware architecture for the preprocessing smoothing filters used in the intra prediction module of the H.266 Versatile Video Coding (H.266/VVC) standard, applicable to both the encoder and decoder sides. The design was fully described in VHDL and synthesized to an Intel Cyclone V FPGA, utilizing 6,961 adaptive logic modules and 4,158 registers, demonstrating feasibility for FPGA implementation with room for integration with a larger intra prediction module. Operating at a maximum frequency of 96.3 MHz without pipelining, the architecture achieves a decoding throughput of up to 2,972 frames per second (fps) for the 1080p resolution, 743 fps for UHD 4K, and 185 fps for UHD 8K, showing that a much lower clock operation can be used for real-time performance. For encoding, the throughput was estimated by targeting a 60 fps rate at 1080p resolution, resulting in an upper bound of approximately 3,171 CUs evaluated per CTU, compatible with fast decision algorithms used in practical encoders.
- Open Access
- 10 November 2025
- Published by Sociedade Brasileira de Computação
- p. 446-454
Traditional vs. Neural Video Codecs: Compression Efficiency, Visual Artifacts, and Quality Analysis Beyond PSNR
Video compression technology is rapidly evolving, with end-to-end learned Neural Video Codecs (NVCs) emerging as powerful alternatives to traditional, block-based standards. This paper presents a comparative analysis of two traditional codecs, HEVC and VVC, against two leading NVCs, DCVC-FM and DCVC-RT. We conduct a thorough rate distortion analysis using multiple objective metrics (PSNR, SSIM, VMAF, and LPIPS) and introduce a formal method to identify rate-matched operating points for fair visual comparison. Our results demonstrate that NVCs offer substantial gains not only in signal fidelity but also in perceptual quality, achieving up to 37.82% bitrate savings over HEVC. Furthermore, our analysis reveals fundamental differences in distortion patterns: NVCs excel at preserving structural and color fidelity, producing visually pleasing results free of blocking artifacts, but can sometimes smooth over fine textures. Conversely, traditional codecs are prone to blockiness but can occasionally retain more high-frequency detail. These findings confirm the superior efficiency of NVCs and highlight the need for evaluation methodologies that look beyond PSNR.
- 10 November 2025
- Published by Sociedade Brasileira de Computação
- p. 446-454
Open AccessTraditional vs. Neural Video Codecs: Compression Efficiency, Visual Artifacts, and Quality Analysis Beyond PSNR
Video compression technology is rapidly evolving, with end-to-end learned Neural Video Codecs (NVCs) emerging as powerful alternatives to traditional, block-based standards. This paper presents a comparative analysis of two traditional codecs, HEVC and VVC, against two leading NVCs, DCVC-FM and DCVC-RT. We conduct a thorough rate distortion analysis using multiple objective metrics (PSNR, SSIM, VMAF, and LPIPS) and introduce a formal method to identify rate-matched operating points for fair visual comparison. Our results demonstrate that NVCs offer substantial gains not only in signal fidelity but also in perceptual quality, achieving up to 37.82% bitrate savings over HEVC. Furthermore, our analysis reveals fundamental differences in distortion patterns: NVCs excel at preserving structural and color fidelity, producing visually pleasing results free of blocking artifacts, but can sometimes smooth over fine textures. Conversely, traditional codecs are prone to blockiness but can occasionally retain more high-frequency detail. These findings confirm the superior efficiency of NVCs and highlight the need for evaluation methodologies that look beyond PSNR.
- Open Access
- 10 November 2025
- Published by Sociedade Brasileira de Computação
- p. 501-505
A Lightweight I3D-Based Approach for Real-Time Brazilian Sign Language Recognition
The demand for accessible technologies to support the deaf and hard-of-hearing community in Brazil is significant. However, many state-of-the-art deep learning models are too computationally intensive for practical, real-time applications. This study addresses this gap by proposing an efficient, lightweight pipeline approach for isolated Brazilian Sign Language (LIBRAS) recognition. We fine-tune a pre-trained Inflated 3D ConvNet (I3D) model on the MINDS-Libras dataset using an end-to-end methodology that operates directly on raw RGB videos, circumventing the need for heavy pre-processing steps like skeleton extraction. To ensure a realistic evaluation of the model’s generalization capabilities, we adopt a strict signer-independent protocol, where test subjects are completely unseen during training. Our proposed model achieves a competitive accuracy of 92.5% and is able to perform sign recognition in real-time, demonstrating strong performance comparable to more complex architectures. This work establishes a new, robust benchmark for signer-agnostic LIBRAS recognition, highlighting that an end-to-end approach can effectively balance high accuracy with the pipeline efficiency required for deployable, real-world accessibility tools.
- 10 November 2025
- Published by Sociedade Brasileira de Computação
- p. 501-505
Open AccessA Lightweight I3D-Based Approach for Real-Time Brazilian Sign Language Recognition
The demand for accessible technologies to support the deaf and hard-of-hearing community in Brazil is significant. However, many state-of-the-art deep learning models are too computationally intensive for practical, real-time applications. This study addresses this gap by proposing an efficient, lightweight pipeline approach for isolated Brazilian Sign Language (LIBRAS) recognition. We fine-tune a pre-trained Inflated 3D ConvNet (I3D) model on the MINDS-Libras dataset using an end-to-end methodology that operates directly on raw RGB videos, circumventing the need for heavy pre-processing steps like skeleton extraction. To ensure a realistic evaluation of the model’s generalization capabilities, we adopt a strict signer-independent protocol, where test subjects are completely unseen during training. Our proposed model achieves a competitive accuracy of 92.5% and is able to perform sign recognition in real-time, demonstrating strong performance comparable to more complex architectures. This work establishes a new, robust benchmark for signer-agnostic LIBRAS recognition, highlighting that an end-to-end approach can effectively balance high accuracy with the pipeline efficiency required for deployable, real-world accessibility tools.